src/intel/vulkan/anv_device.c

   1 /*
   2  * Copyright © 2015 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <assert.h>
  25 #include <stdbool.h>
  26 #include <string.h>
  27 #include <sys/mman.h>
  28 #include <unistd.h>
  29 #include <fcntl.h>
  30
  31 #include "anv_private.h"
  32 #include "util/strtod.h"
  33 #include "util/debug.h"
  34 #include "util/build_id.h"
  35 #include "util/vk_util.h"
  36
  37 #include "genxml/gen7_pack.h"
  38
  39 static void
  40 compiler_debug_log(void *data, const char *fmt, ...)
  41 { }
  42
  43 static void
  44 compiler_perf_log(void *data, const char *fmt, ...)
  45 {
  46    va_list args;
  47    va_start(args, fmt);
  48
  49    if (unlikely(INTEL_DEBUG & DEBUG_PERF))
  50       vfprintf(stderr, fmt, args);
  51
  52    va_end(args);
  53 }
  54
  55 static bool
  56 anv_device_get_cache_uuid(void *uuid)
  57 {
  58    const struct build_id_note *note = build_id_find_nhdr("libvulkan_intel.so");
  59    if (!note)
  60       return false;
  61
  62    unsigned len = build_id_length(note);
  63    if (len < VK_UUID_SIZE)
  64       return false;
  65
  66    memcpy(uuid, build_id_data(note), VK_UUID_SIZE);
  67    return true;
  68 }
  69
  70 static VkResult
  71 anv_physical_device_init(struct anv_physical_device *device,
  72                          struct anv_instance *instance,
  73                          const char *path)
  74 {
  75    VkResult result;
  76    int fd;
  77
  78    fd = open(path, O_RDWR | O_CLOEXEC);
  79    if (fd < 0)
  80       return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  81
  82    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
  83    device->instance = instance;
  84
  85    assert(strlen(path) < ARRAY_SIZE(device->path));
  86    strncpy(device->path, path, ARRAY_SIZE(device->path));
  87
  88    device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
  89    if (!device->chipset_id) {
  90       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  91       goto fail;
  92    }
  93
  94    device->name = gen_get_device_name(device->chipset_id);
  95    if (!gen_get_device_info(device->chipset_id, &device->info)) {
  96       result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
  97       goto fail;
  98    }
  99
 100    if (device->info.is_haswell) {
 101       fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
 102    } else if (device->info.gen == 7 && !device->info.is_baytrail) {
 103       fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
 104    } else if (device->info.gen == 7 && device->info.is_baytrail) {
 105       fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
 106    } else if (device->info.gen >= 8) {
 107       /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
 108        * supported as anything */
 109    } else {
 110       result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 111                          "Vulkan not yet supported on %s", device->name);
 112       goto fail;
 113    }
 114
 115    device->cmd_parser_version = -1;
 116    if (device->info.gen == 7) {
 117       device->cmd_parser_version =
 118          anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
 119       if (device->cmd_parser_version == -1) {
 120          result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 121                             "failed to get command parser version");
 122          goto fail;
 123       }
 124    }
 125
 126    if (anv_gem_get_aperture(fd, &device->aperture_size) == -1) {
 127       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 128                          "failed to get aperture size: %m");
 129       goto fail;
 130    }
 131
 132    if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
 133       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 134                          "kernel missing gem wait");
 135       goto fail;
 136    }
 137
 138    if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
 139       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 140                          "kernel missing execbuf2");
 141       goto fail;
 142    }
 143
 144    if (!device->info.has_llc &&
 145        anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
 146       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 147                          "kernel missing wc mmap");
 148       goto fail;
 149    }
 150
 151    if (!anv_device_get_cache_uuid(device->uuid)) {
 152       result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
 153                          "cannot generate UUID");
 154       goto fail;
 155    }
 156    bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
 157
 158    /* GENs prior to 8 do not support EU/Subslice info */
 159    if (device->info.gen >= 8) {
 160       device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
 161       device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
 162
 163       /* Without this information, we cannot get the right Braswell
 164        * brandstrings, and we have to use conservative numbers for GPGPU on
 165        * many platforms, but otherwise, things will just work.
 166        */
 167       if (device->subslice_total < 1 || device->eu_total < 1) {
 168          fprintf(stderr, "WARNING: Kernel 4.1 required to properly"
 169                          " query GPU properties.\n");
 170       }
 171    } else if (device->info.gen == 7) {
 172       device->subslice_total = 1 << (device->info.gt - 1);
 173    }
 174
 175    if (device->info.is_cherryview &&
 176        device->subslice_total > 0 && device->eu_total > 0) {
 177       /* Logical CS threads = EUs per subslice * 7 threads per EU */
 178       uint32_t max_cs_threads = device->eu_total / device->subslice_total * 7;
 179
 180       /* Fuse configurations may give more threads than expected, never less. */
 181       if (max_cs_threads > device->info.max_cs_threads)
 182          device->info.max_cs_threads = max_cs_threads;
 183    }
 184
 185    brw_process_intel_debug_variable();
 186
 187    device->compiler = brw_compiler_create(NULL, &device->info);
 188    if (device->compiler == NULL) {
 189       result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 190       goto fail;
 191    }
 192    device->compiler->shader_debug_log = compiler_debug_log;
 193    device->compiler->shader_perf_log = compiler_perf_log;
 194
 195    result = anv_init_wsi(device);
 196    if (result != VK_SUCCESS) {
 197       ralloc_free(device->compiler);
 198       goto fail;
 199    }
 200
 201    isl_device_init(&device->isl_dev, &device->info, swizzled);
 202
 203    device->local_fd = fd;
 204    return VK_SUCCESS;
 205
 206 fail:
 207    close(fd);
 208    return result;
 209 }
 210
 211 static void
 212 anv_physical_device_finish(struct anv_physical_device *device)
 213 {
 214    anv_finish_wsi(device);
 215    ralloc_free(device->compiler);
 216    close(device->local_fd);
 217 }
 218
 219 static const VkExtensionProperties global_extensions[] = {
 220    {
 221       .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
 222       .specVersion = 25,
 223    },
 224 #ifdef VK_USE_PLATFORM_XCB_KHR
 225    {
 226       .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
 227       .specVersion = 6,
 228    },
 229 #endif
 230 #ifdef VK_USE_PLATFORM_XLIB_KHR
 231    {
 232       .extensionName = VK_KHR_XLIB_SURFACE_EXTENSION_NAME,
 233       .specVersion = 6,
 234    },
 235 #endif
 236 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
 237    {
 238       .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
 239       .specVersion = 5,
 240    },
 241 #endif
 242    {
 243       .extensionName = VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
 244       .specVersion = 1,
 245    },
 246 };
 247
 248 static const VkExtensionProperties device_extensions[] = {
 249    {
 250       .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
 251       .specVersion = 68,
 252    },
 253    {
 254       .extensionName = VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
 255       .specVersion = 1,
 256    },
 257    {
 258       .extensionName = VK_KHR_MAINTENANCE1_EXTENSION_NAME,
 259       .specVersion = 1,
 260    },
 261    {
 262       .extensionName = VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
 263       .specVersion = 1,
 264    },
 265    {
 266       .extensionName = VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME,
 267       .specVersion = 1,
 268    }
 269 };
 270
 271 static void *
 272 default_alloc_func(void *pUserData, size_t size, size_t align,
 273                    VkSystemAllocationScope allocationScope)
 274 {
 275    return malloc(size);
 276 }
 277
 278 static void *
 279 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
 280                      size_t align, VkSystemAllocationScope allocationScope)
 281 {
 282    return realloc(pOriginal, size);
 283 }
 284
 285 static void
 286 default_free_func(void *pUserData, void *pMemory)
 287 {
 288    free(pMemory);
 289 }
 290
 291 static const VkAllocationCallbacks default_alloc = {
 292    .pUserData = NULL,
 293    .pfnAllocation = default_alloc_func,
 294    .pfnReallocation = default_realloc_func,
 295    .pfnFree = default_free_func,
 296 };
 297
 298 VkResult anv_CreateInstance(
 299     const VkInstanceCreateInfo*                 pCreateInfo,
 300     const VkAllocationCallbacks*                pAllocator,
 301     VkInstance*                                 pInstance)
 302 {
 303    struct anv_instance *instance;
 304
 305    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 306
 307    uint32_t client_version;
 308    if (pCreateInfo->pApplicationInfo &&
 309        pCreateInfo->pApplicationInfo->apiVersion != 0) {
 310       client_version = pCreateInfo->pApplicationInfo->apiVersion;
 311    } else {
 312       client_version = VK_MAKE_VERSION(1, 0, 0);
 313    }
 314
 315    if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
 316        client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
 317       return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
 318                        "Client requested version %d.%d.%d",
 319                        VK_VERSION_MAJOR(client_version),
 320                        VK_VERSION_MINOR(client_version),
 321                        VK_VERSION_PATCH(client_version));
 322    }
 323
 324    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 325       bool found = false;
 326       for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
 327          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 328                     global_extensions[j].extensionName) == 0) {
 329             found = true;
 330             break;
 331          }
 332       }
 333       if (!found)
 334          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 335    }
 336
 337    instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
 338                          VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
 339    if (!instance)
 340       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 341
 342    instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 343
 344    if (pAllocator)
 345       instance->alloc = *pAllocator;
 346    else
 347       instance->alloc = default_alloc;
 348
 349    instance->apiVersion = client_version;
 350    instance->physicalDeviceCount = -1;
 351
 352    _mesa_locale_init();
 353
 354    VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
 355
 356    *pInstance = anv_instance_to_handle(instance);
 357
 358    return VK_SUCCESS;
 359 }
 360
 361 void anv_DestroyInstance(
 362     VkInstance                                  _instance,
 363     const VkAllocationCallbacks*                pAllocator)
 364 {
 365    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 366
 367    if (!instance)
 368       return;
 369
 370    if (instance->physicalDeviceCount > 0) {
 371       /* We support at most one physical device. */
 372       assert(instance->physicalDeviceCount == 1);
 373       anv_physical_device_finish(&instance->physicalDevice);
 374    }
 375
 376    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 377
 378    _mesa_locale_fini();
 379
 380    vk_free(&instance->alloc, instance);
 381 }
 382
 383 VkResult anv_EnumeratePhysicalDevices(
 384     VkInstance                                  _instance,
 385     uint32_t*                                   pPhysicalDeviceCount,
 386     VkPhysicalDevice*                           pPhysicalDevices)
 387 {
 388    ANV_FROM_HANDLE(anv_instance, instance, _instance);
 389    VkResult result;
 390
 391    if (instance->physicalDeviceCount < 0) {
 392       char path[20];
 393       for (unsigned i = 0; i < 8; i++) {
 394          snprintf(path, sizeof(path), "/dev/dri/renderD%d", 128 + i);
 395          result = anv_physical_device_init(&instance->physicalDevice,
 396                                            instance, path);
 397          if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
 398             break;
 399       }
 400
 401       if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
 402          instance->physicalDeviceCount = 0;
 403       } else if (result == VK_SUCCESS) {
 404          instance->physicalDeviceCount = 1;
 405       } else {
 406          return result;
 407       }
 408    }
 409
 410    /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
 411     * otherwise it's an inout parameter.
 412     *
 413     * The Vulkan spec (git aaed022) says:
 414     *
 415     *    pPhysicalDeviceCount is a pointer to an unsigned integer variable
 416     *    that is initialized with the number of devices the application is
 417     *    prepared to receive handles to. pname:pPhysicalDevices is pointer to
 418     *    an array of at least this many VkPhysicalDevice handles [...].
 419     *
 420     *    Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
 421     *    overwrites the contents of the variable pointed to by
 422     *    pPhysicalDeviceCount with the number of physical devices in in the
 423     *    instance; otherwise, vkEnumeratePhysicalDevices overwrites
 424     *    pPhysicalDeviceCount with the number of physical handles written to
 425     *    pPhysicalDevices.
 426     */
 427    if (!pPhysicalDevices) {
 428       *pPhysicalDeviceCount = instance->physicalDeviceCount;
 429    } else if (*pPhysicalDeviceCount >= 1) {
 430       pPhysicalDevices[0] = anv_physical_device_to_handle(&instance->physicalDevice);
 431       *pPhysicalDeviceCount = 1;
 432    } else if (*pPhysicalDeviceCount < instance->physicalDeviceCount) {
 433       return VK_INCOMPLETE;
 434    } else {
 435       *pPhysicalDeviceCount = 0;
 436    }
 437
 438    return VK_SUCCESS;
 439 }
 440
 441 void anv_GetPhysicalDeviceFeatures(
 442     VkPhysicalDevice                            physicalDevice,
 443     VkPhysicalDeviceFeatures*                   pFeatures)
 444 {
 445    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 446
 447    *pFeatures = (VkPhysicalDeviceFeatures) {
 448       .robustBufferAccess                       = true,
 449       .fullDrawIndexUint32                      = true,
 450       .imageCubeArray                           = true,
 451       .independentBlend                         = true,
 452       .geometryShader                           = true,
 453       .tessellationShader                       = true,
 454       .sampleRateShading                        = true,
 455       .dualSrcBlend                             = true,
 456       .logicOp                                  = true,
 457       .multiDrawIndirect                        = false,
 458       .drawIndirectFirstInstance                = true,
 459       .depthClamp                               = true,
 460       .depthBiasClamp                           = true,
 461       .fillModeNonSolid                         = true,
 462       .depthBounds                              = false,
 463       .wideLines                                = true,
 464       .largePoints                              = true,
 465       .alphaToOne                               = true,
 466       .multiViewport                            = true,
 467       .samplerAnisotropy                        = true,
 468       .textureCompressionETC2                   = pdevice->info.gen >= 8 ||
 469                                                   pdevice->info.is_baytrail,
 470       .textureCompressionASTC_LDR               = pdevice->info.gen >= 9, /* FINISHME CHV */
 471       .textureCompressionBC                     = true,
 472       .occlusionQueryPrecise                    = true,
 473       .pipelineStatisticsQuery                  = false,
 474       .fragmentStoresAndAtomics                 = true,
 475       .shaderTessellationAndGeometryPointSize   = true,
 476       .shaderImageGatherExtended                = true,
 477       .shaderStorageImageExtendedFormats        = true,
 478       .shaderStorageImageMultisample            = false,
 479       .shaderStorageImageReadWithoutFormat      = false,
 480       .shaderStorageImageWriteWithoutFormat     = true,
 481       .shaderUniformBufferArrayDynamicIndexing  = true,
 482       .shaderSampledImageArrayDynamicIndexing   = true,
 483       .shaderStorageBufferArrayDynamicIndexing  = true,
 484       .shaderStorageImageArrayDynamicIndexing   = true,
 485       .shaderClipDistance                       = true,
 486       .shaderCullDistance                       = true,
 487       .shaderFloat64                            = pdevice->info.gen >= 8,
 488       .shaderInt64                              = false,
 489       .shaderInt16                              = false,
 490       .shaderResourceMinLod                     = false,
 491       .variableMultisampleRate                  = false,
 492       .inheritedQueries                         = false,
 493    };
 494
 495    /* We can't do image stores in vec4 shaders */
 496    pFeatures->vertexPipelineStoresAndAtomics =
 497       pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
 498       pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
 499 }
 500
 501 void anv_GetPhysicalDeviceFeatures2KHR(
 502     VkPhysicalDevice                            physicalDevice,
 503     VkPhysicalDeviceFeatures2KHR*               pFeatures)
 504 {
 505    anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 506
 507    vk_foreach_struct(ext, pFeatures->pNext) {
 508       switch (ext->sType) {
 509       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
 510          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
 511             (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
 512
 513          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
 514          break;
 515       }
 516
 517       default:
 518          anv_debug_ignored_stype(ext->sType);
 519          break;
 520       }
 521    }
 522 }
 523
 524 void anv_GetPhysicalDeviceProperties(
 525     VkPhysicalDevice                            physicalDevice,
 526     VkPhysicalDeviceProperties*                 pProperties)
 527 {
 528    ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
 529    const struct gen_device_info *devinfo = &pdevice->info;
 530
 531    const float time_stamp_base = devinfo->gen >= 9 ? 83.333 : 80.0;
 532
 533    /* See assertions made when programming the buffer surface state. */
 534    const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
 535                                       (1ul << 30) : (1ul << 27);
 536
 537    VkSampleCountFlags sample_counts =
 538       isl_device_get_sample_counts(&pdevice->isl_dev);
 539
 540    VkPhysicalDeviceLimits limits = {
 541       .maxImageDimension1D                      = (1 << 14),
 542       .maxImageDimension2D                      = (1 << 14),
 543       .maxImageDimension3D                      = (1 << 11),
 544       .maxImageDimensionCube                    = (1 << 14),
 545       .maxImageArrayLayers                      = (1 << 11),
 546       .maxTexelBufferElements                   = 128 * 1024 * 1024,
 547       .maxUniformBufferRange                    = (1ul << 27),
 548       .maxStorageBufferRange                    = max_raw_buffer_sz,
 549       .maxPushConstantsSize                     = MAX_PUSH_CONSTANTS_SIZE,
 550       .maxMemoryAllocationCount                 = UINT32_MAX,
 551       .maxSamplerAllocationCount                = 64 * 1024,
 552       .bufferImageGranularity                   = 64, /* A cache line */
 553       .sparseAddressSpaceSize                   = 0,
 554       .maxBoundDescriptorSets                   = MAX_SETS,
 555       .maxPerStageDescriptorSamplers            = 64,
 556       .maxPerStageDescriptorUniformBuffers      = 64,
 557       .maxPerStageDescriptorStorageBuffers      = 64,
 558       .maxPerStageDescriptorSampledImages       = 64,
 559       .maxPerStageDescriptorStorageImages       = 64,
 560       .maxPerStageDescriptorInputAttachments    = 64,
 561       .maxPerStageResources                     = 128,
 562       .maxDescriptorSetSamplers                 = 256,
 563       .maxDescriptorSetUniformBuffers           = 256,
 564       .maxDescriptorSetUniformBuffersDynamic    = 256,
 565       .maxDescriptorSetStorageBuffers           = 256,
 566       .maxDescriptorSetStorageBuffersDynamic    = 256,
 567       .maxDescriptorSetSampledImages            = 256,
 568       .maxDescriptorSetStorageImages            = 256,
 569       .maxDescriptorSetInputAttachments         = 256,
 570       .maxVertexInputAttributes                 = MAX_VBS,
 571       .maxVertexInputBindings                   = MAX_VBS,
 572       .maxVertexInputAttributeOffset            = 2047,
 573       .maxVertexInputBindingStride              = 2048,
 574       .maxVertexOutputComponents                = 128,
 575       .maxTessellationGenerationLevel           = 64,
 576       .maxTessellationPatchSize                 = 32,
 577       .maxTessellationControlPerVertexInputComponents = 128,
 578       .maxTessellationControlPerVertexOutputComponents = 128,
 579       .maxTessellationControlPerPatchOutputComponents = 128,
 580       .maxTessellationControlTotalOutputComponents = 2048,
 581       .maxTessellationEvaluationInputComponents = 128,
 582       .maxTessellationEvaluationOutputComponents = 128,
 583       .maxGeometryShaderInvocations             = 32,
 584       .maxGeometryInputComponents               = 64,
 585       .maxGeometryOutputComponents              = 128,
 586       .maxGeometryOutputVertices                = 256,
 587       .maxGeometryTotalOutputComponents         = 1024,
 588       .maxFragmentInputComponents               = 128,
 589       .maxFragmentOutputAttachments             = 8,
 590       .maxFragmentDualSrcAttachments            = 1,
 591       .maxFragmentCombinedOutputResources       = 8,
 592       .maxComputeSharedMemorySize               = 32768,
 593       .maxComputeWorkGroupCount                 = { 65535, 65535, 65535 },
 594       .maxComputeWorkGroupInvocations           = 16 * devinfo->max_cs_threads,
 595       .maxComputeWorkGroupSize = {
 596          16 * devinfo->max_cs_threads,
 597          16 * devinfo->max_cs_threads,
 598          16 * devinfo->max_cs_threads,
 599       },
 600       .subPixelPrecisionBits                    = 4 /* FIXME */,
 601       .subTexelPrecisionBits                    = 4 /* FIXME */,
 602       .mipmapPrecisionBits                      = 4 /* FIXME */,
 603       .maxDrawIndexedIndexValue                 = UINT32_MAX,
 604       .maxDrawIndirectCount                     = UINT32_MAX,
 605       .maxSamplerLodBias                        = 16,
 606       .maxSamplerAnisotropy                     = 16,
 607       .maxViewports                             = MAX_VIEWPORTS,
 608       .maxViewportDimensions                    = { (1 << 14), (1 << 14) },
 609       .viewportBoundsRange                      = { INT16_MIN, INT16_MAX },
 610       .viewportSubPixelBits                     = 13, /* We take a float? */
 611       .minMemoryMapAlignment                    = 4096, /* A page */
 612       .minTexelBufferOffsetAlignment            = 1,
 613       .minUniformBufferOffsetAlignment          = 16,
 614       .minStorageBufferOffsetAlignment          = 4,
 615       .minTexelOffset                           = -8,
 616       .maxTexelOffset                           = 7,
 617       .minTexelGatherOffset                     = -32,
 618       .maxTexelGatherOffset                     = 31,
 619       .minInterpolationOffset                   = -0.5,
 620       .maxInterpolationOffset                   = 0.4375,
 621       .subPixelInterpolationOffsetBits          = 4,
 622       .maxFramebufferWidth                      = (1 << 14),
 623       .maxFramebufferHeight                     = (1 << 14),
 624       .maxFramebufferLayers                     = (1 << 11),
 625       .framebufferColorSampleCounts             = sample_counts,
 626       .framebufferDepthSampleCounts             = sample_counts,
 627       .framebufferStencilSampleCounts           = sample_counts,
 628       .framebufferNoAttachmentsSampleCounts     = sample_counts,
 629       .maxColorAttachments                      = MAX_RTS,
 630       .sampledImageColorSampleCounts            = sample_counts,
 631       .sampledImageIntegerSampleCounts          = VK_SAMPLE_COUNT_1_BIT,
 632       .sampledImageDepthSampleCounts            = sample_counts,
 633       .sampledImageStencilSampleCounts          = sample_counts,
 634       .storageImageSampleCounts                 = VK_SAMPLE_COUNT_1_BIT,
 635       .maxSampleMaskWords                       = 1,
 636       .timestampComputeAndGraphics              = false,
 637       .timestampPeriod                          = time_stamp_base,
 638       .maxClipDistances                         = 8,
 639       .maxCullDistances                         = 8,
 640       .maxCombinedClipAndCullDistances          = 8,
 641       .discreteQueuePriorities                  = 1,
 642       .pointSizeRange                           = { 0.125, 255.875 },
 643       .lineWidthRange                           = { 0.0, 7.9921875 },
 644       .pointSizeGranularity                     = (1.0 / 8.0),
 645       .lineWidthGranularity                     = (1.0 / 128.0),
 646       .strictLines                              = false, /* FINISHME */
 647       .standardSampleLocations                  = true,
 648       .optimalBufferCopyOffsetAlignment         = 128,
 649       .optimalBufferCopyRowPitchAlignment       = 128,
 650       .nonCoherentAtomSize                      = 64,
 651    };
 652
 653    *pProperties = (VkPhysicalDeviceProperties) {
 654       .apiVersion = VK_MAKE_VERSION(1, 0, 42),
 655       .driverVersion = 1,
 656       .vendorID = 0x8086,
 657       .deviceID = pdevice->chipset_id,
 658       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
 659       .limits = limits,
 660       .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
 661    };
 662
 663    strcpy(pProperties->deviceName, pdevice->name);
 664    memcpy(pProperties->pipelineCacheUUID, pdevice->uuid, VK_UUID_SIZE);
 665 }
 666
 667 void anv_GetPhysicalDeviceProperties2KHR(
 668     VkPhysicalDevice                            physicalDevice,
 669     VkPhysicalDeviceProperties2KHR*             pProperties)
 670 {
 671    anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 672
 673    vk_foreach_struct(ext, pProperties->pNext) {
 674       switch (ext->sType) {
 675       default:
 676          anv_debug_ignored_stype(ext->sType);
 677          break;
 678       }
 679    }
 680 }
 681
 682 static void
 683 anv_get_queue_family_properties(struct anv_physical_device *phys_dev,
 684                                 VkQueueFamilyProperties *props)
 685 {
 686    *props = (VkQueueFamilyProperties) {
 687       .queueFlags = VK_QUEUE_GRAPHICS_BIT |
 688                     VK_QUEUE_COMPUTE_BIT |
 689                     VK_QUEUE_TRANSFER_BIT,
 690       .queueCount = 1,
 691       .timestampValidBits = 36, /* XXX: Real value here */
 692       .minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
 693    };
 694 }
 695
 696 void anv_GetPhysicalDeviceQueueFamilyProperties(
 697     VkPhysicalDevice                            physicalDevice,
 698     uint32_t*                                   pCount,
 699     VkQueueFamilyProperties*                    pQueueFamilyProperties)
 700 {
 701    ANV_FROM_HANDLE(anv_physical_device, phys_dev, physicalDevice);
 702
 703    if (pQueueFamilyProperties == NULL) {
 704       *pCount = 1;
 705       return;
 706    }
 707
 708    /* The spec implicitly allows the incoming count to be 0. From the Vulkan
 709     * 1.0.38 spec, Section 4.1 Physical Devices:
 710     *
 711     *     If the value referenced by pQueueFamilyPropertyCount is not 0 [then
 712     *     do stuff].
 713     */
 714    if (*pCount == 0)
 715       return;
 716
 717    *pCount = 1;
 718    anv_get_queue_family_properties(phys_dev, pQueueFamilyProperties);
 719 }
 720
 721 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
 722     VkPhysicalDevice                            physicalDevice,
 723     uint32_t*                                   pQueueFamilyPropertyCount,
 724     VkQueueFamilyProperties2KHR*                pQueueFamilyProperties)
 725 {
 726
 727    ANV_FROM_HANDLE(anv_physical_device, phys_dev, physicalDevice);
 728
 729    if (pQueueFamilyProperties == NULL) {
 730       *pQueueFamilyPropertyCount = 1;
 731       return;
 732    }
 733
 734    /* The spec implicitly allows the incoming count to be 0. From the Vulkan
 735     * 1.0.38 spec, Section 4.1 Physical Devices:
 736     *
 737     *     If the value referenced by pQueueFamilyPropertyCount is not 0 [then
 738     *     do stuff].
 739     */
 740    if (*pQueueFamilyPropertyCount == 0)
 741       return;
 742
 743    /* We support exactly one queue family. So need to traverse only the first
 744     * array element's pNext chain.
 745     */
 746    *pQueueFamilyPropertyCount = 1;
 747    anv_get_queue_family_properties(phys_dev,
 748          &pQueueFamilyProperties->queueFamilyProperties);
 749
 750    vk_foreach_struct(ext, pQueueFamilyProperties->pNext) {
 751       switch (ext->sType) {
 752       default:
 753          anv_debug_ignored_stype(ext->sType);
 754          break;
 755       }
 756    }
 757 }
 758
 759 void anv_GetPhysicalDeviceMemoryProperties(
 760     VkPhysicalDevice                            physicalDevice,
 761     VkPhysicalDeviceMemoryProperties*           pMemoryProperties)
 762 {
 763    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 764    VkDeviceSize heap_size;
 765
 766    /* Reserve some wiggle room for the driver by exposing only 75% of the
 767     * aperture to the heap.
 768     */
 769    heap_size = 3 * physical_device->aperture_size / 4;
 770
 771    if (physical_device->info.has_llc) {
 772       /* Big core GPUs share LLC with the CPU and thus one memory type can be
 773        * both cached and coherent at the same time.
 774        */
 775       pMemoryProperties->memoryTypeCount = 1;
 776       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 777          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 778                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 779                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
 780                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 781          .heapIndex = 0,
 782       };
 783    } else {
 784       /* The spec requires that we expose a host-visible, coherent memory
 785        * type, but Atom GPUs don't share LLC. Thus we offer two memory types
 786        * to give the application a choice between cached, but not coherent and
 787        * coherent but uncached (WC though).
 788        */
 789       pMemoryProperties->memoryTypeCount = 2;
 790       pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
 791          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 792                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 793                           VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
 794          .heapIndex = 0,
 795       };
 796       pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
 797          .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
 798                           VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
 799                           VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
 800          .heapIndex = 0,
 801       };
 802    }
 803
 804    pMemoryProperties->memoryHeapCount = 1;
 805    pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
 806       .size = heap_size,
 807       .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
 808    };
 809 }
 810
 811 void anv_GetPhysicalDeviceMemoryProperties2KHR(
 812     VkPhysicalDevice                            physicalDevice,
 813     VkPhysicalDeviceMemoryProperties2KHR*       pMemoryProperties)
 814 {
 815    anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
 816                                          &pMemoryProperties->memoryProperties);
 817
 818    vk_foreach_struct(ext, pMemoryProperties->pNext) {
 819       switch (ext->sType) {
 820       default:
 821          anv_debug_ignored_stype(ext->sType);
 822          break;
 823       }
 824    }
 825 }
 826
 827 PFN_vkVoidFunction anv_GetInstanceProcAddr(
 828     VkInstance                                  instance,
 829     const char*                                 pName)
 830 {
 831    return anv_lookup_entrypoint(NULL, pName);
 832 }
 833
 834 /* With version 1+ of the loader interface the ICD should expose
 835  * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
 836  */
 837 PUBLIC
 838 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 839     VkInstance                                  instance,
 840     const char*                                 pName);
 841
 842 PUBLIC
 843 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
 844     VkInstance                                  instance,
 845     const char*                                 pName)
 846 {
 847    return anv_GetInstanceProcAddr(instance, pName);
 848 }
 849
 850 PFN_vkVoidFunction anv_GetDeviceProcAddr(
 851     VkDevice                                    _device,
 852     const char*                                 pName)
 853 {
 854    ANV_FROM_HANDLE(anv_device, device, _device);
 855    return anv_lookup_entrypoint(&device->info, pName);
 856 }
 857
 858 static void
 859 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
 860 {
 861    queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 862    queue->device = device;
 863    queue->pool = &device->surface_state_pool;
 864 }
 865
 866 static void
 867 anv_queue_finish(struct anv_queue *queue)
 868 {
 869 }
 870
 871 static struct anv_state
 872 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
 873 {
 874    struct anv_state state;
 875
 876    state = anv_state_pool_alloc(pool, size, align);
 877    memcpy(state.map, p, size);
 878
 879    anv_state_flush(pool->block_pool->device, state);
 880
 881    return state;
 882 }
 883
 884 struct gen8_border_color {
 885    union {
 886       float float32[4];
 887       uint32_t uint32[4];
 888    };
 889    /* Pad out to 64 bytes */
 890    uint32_t _pad[12];
 891 };
 892
 893 static void
 894 anv_device_init_border_colors(struct anv_device *device)
 895 {
 896    static const struct gen8_border_color border_colors[] = {
 897       [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] =  { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
 898       [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] =       { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
 899       [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] =       { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
 900       [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] =    { .uint32 = { 0, 0, 0, 0 } },
 901       [VK_BORDER_COLOR_INT_OPAQUE_BLACK] =         { .uint32 = { 0, 0, 0, 1 } },
 902       [VK_BORDER_COLOR_INT_OPAQUE_WHITE] =         { .uint32 = { 1, 1, 1, 1 } },
 903    };
 904
 905    device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
 906                                                     sizeof(border_colors), 64,
 907                                                     border_colors);
 908 }
 909
 910 VkResult
 911 anv_device_submit_simple_batch(struct anv_device *device,
 912                                struct anv_batch *batch)
 913 {
 914    struct drm_i915_gem_execbuffer2 execbuf;
 915    struct drm_i915_gem_exec_object2 exec2_objects[1];
 916    struct anv_bo bo, *exec_bos[1];
 917    VkResult result = VK_SUCCESS;
 918    uint32_t size;
 919    int64_t timeout;
 920    int ret;
 921
 922    /* Kernel driver requires 8 byte aligned batch length */
 923    size = align_u32(batch->next - batch->start, 8);
 924    result = anv_bo_pool_alloc(&device->batch_bo_pool, &bo, size);
 925    if (result != VK_SUCCESS)
 926       return result;
 927
 928    memcpy(bo.map, batch->start, size);
 929    if (!device->info.has_llc)
 930       anv_flush_range(bo.map, size);
 931
 932    exec_bos[0] = &bo;
 933    exec2_objects[0].handle = bo.gem_handle;
 934    exec2_objects[0].relocation_count = 0;
 935    exec2_objects[0].relocs_ptr = 0;
 936    exec2_objects[0].alignment = 0;
 937    exec2_objects[0].offset = bo.offset;
 938    exec2_objects[0].flags = 0;
 939    exec2_objects[0].rsvd1 = 0;
 940    exec2_objects[0].rsvd2 = 0;
 941
 942    execbuf.buffers_ptr = (uintptr_t) exec2_objects;
 943    execbuf.buffer_count = 1;
 944    execbuf.batch_start_offset = 0;
 945    execbuf.batch_len = size;
 946    execbuf.cliprects_ptr = 0;
 947    execbuf.num_cliprects = 0;
 948    execbuf.DR1 = 0;
 949    execbuf.DR4 = 0;
 950
 951    execbuf.flags =
 952       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
 953    execbuf.rsvd1 = device->context_id;
 954    execbuf.rsvd2 = 0;
 955
 956    result = anv_device_execbuf(device, &execbuf, exec_bos);
 957    if (result != VK_SUCCESS)
 958       goto fail;
 959
 960    timeout = INT64_MAX;
 961    ret = anv_gem_wait(device, bo.gem_handle, &timeout);
 962    if (ret != 0) {
 963       /* We don't know the real error. */
 964       result = vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
 965       goto fail;
 966    }
 967
 968  fail:
 969    anv_bo_pool_free(&device->batch_bo_pool, &bo);
 970
 971    return result;
 972 }
 973
 974 VkResult anv_CreateDevice(
 975     VkPhysicalDevice                            physicalDevice,
 976     const VkDeviceCreateInfo*                   pCreateInfo,
 977     const VkAllocationCallbacks*                pAllocator,
 978     VkDevice*                                   pDevice)
 979 {
 980    ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
 981    VkResult result;
 982    struct anv_device *device;
 983
 984    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
 985
 986    for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
 987       bool found = false;
 988       for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
 989          if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
 990                     device_extensions[j].extensionName) == 0) {
 991             found = true;
 992             break;
 993          }
 994       }
 995       if (!found)
 996          return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
 997    }
 998
 999    device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
1000                        sizeof(*device), 8,
1001                        VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
1002    if (!device)
1003       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1004
1005    device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1006    device->instance = physical_device->instance;
1007    device->chipset_id = physical_device->chipset_id;
1008
1009    if (pAllocator)
1010       device->alloc = *pAllocator;
1011    else
1012       device->alloc = physical_device->instance->alloc;
1013
1014    /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1015    device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
1016    if (device->fd == -1) {
1017       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1018       goto fail_device;
1019    }
1020
1021    device->context_id = anv_gem_create_context(device);
1022    if (device->context_id == -1) {
1023       result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
1024       goto fail_fd;
1025    }
1026
1027    device->info = physical_device->info;
1028    device->isl_dev = physical_device->isl_dev;
1029
1030    /* On Broadwell and later, we can use batch chaining to more efficiently
1031     * implement growing command buffers.  Prior to Haswell, the kernel
1032     * command parser gets in the way and we have to fall back to growing
1033     * the batch.
1034     */
1035    device->can_chain_batches = device->info.gen >= 8;
1036
1037    device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
1038       pCreateInfo->pEnabledFeatures->robustBufferAccess;
1039
1040    pthread_mutex_init(&device->mutex, NULL);
1041
1042    pthread_condattr_t condattr;
1043    pthread_condattr_init(&condattr);
1044    pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC);
1045    pthread_cond_init(&device->queue_submit, NULL);
1046    pthread_condattr_destroy(&condattr);
1047
1048    anv_bo_pool_init(&device->batch_bo_pool, device);
1049
1050    anv_block_pool_init(&device->dynamic_state_block_pool, device, 16384);
1051
1052    anv_state_pool_init(&device->dynamic_state_pool,
1053                        &device->dynamic_state_block_pool);
1054
1055    anv_block_pool_init(&device->instruction_block_pool, device, 1024 * 1024);
1056    anv_state_pool_init(&device->instruction_state_pool,
1057                        &device->instruction_block_pool);
1058
1059    anv_block_pool_init(&device->surface_state_block_pool, device, 4096);
1060
1061    anv_state_pool_init(&device->surface_state_pool,
1062                        &device->surface_state_block_pool);
1063
1064    anv_bo_init_new(&device->workaround_bo, device, 1024);
1065
1066    anv_scratch_pool_init(device, &device->scratch_pool);
1067
1068    anv_queue_init(device, &device->queue);
1069
1070    switch (device->info.gen) {
1071    case 7:
1072       if (!device->info.is_haswell)
1073          result = gen7_init_device_state(device);
1074       else
1075          result = gen75_init_device_state(device);
1076       break;
1077    case 8:
1078       result = gen8_init_device_state(device);
1079       break;
1080    case 9:
1081       result = gen9_init_device_state(device);
1082       break;
1083    default:
1084       /* Shouldn't get here as we don't create physical devices for any other
1085        * gens. */
1086       unreachable("unhandled gen");
1087    }
1088    if (result != VK_SUCCESS)
1089       goto fail_fd;
1090
1091    anv_device_init_blorp(device);
1092
1093    anv_device_init_border_colors(device);
1094
1095    *pDevice = anv_device_to_handle(device);
1096
1097    return VK_SUCCESS;
1098
1099  fail_fd:
1100    close(device->fd);
1101  fail_device:
1102    vk_free(&device->alloc, device);
1103
1104    return result;
1105 }
1106
1107 void anv_DestroyDevice(
1108     VkDevice                                    _device,
1109     const VkAllocationCallbacks*                pAllocator)
1110 {
1111    ANV_FROM_HANDLE(anv_device, device, _device);
1112
1113    if (!device)
1114       return;
1115
1116    anv_device_finish_blorp(device);
1117
1118    anv_queue_finish(&device->queue);
1119
1120 #ifdef HAVE_VALGRIND
1121    /* We only need to free these to prevent valgrind errors.  The backing
1122     * BO will go away in a couple of lines so we don't actually leak.
1123     */
1124    anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
1125 #endif
1126
1127    anv_scratch_pool_finish(device, &device->scratch_pool);
1128
1129    anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1130    anv_gem_close(device, device->workaround_bo.gem_handle);
1131
1132    anv_state_pool_finish(&device->surface_state_pool);
1133    anv_block_pool_finish(&device->surface_state_block_pool);
1134    anv_state_pool_finish(&device->instruction_state_pool);
1135    anv_block_pool_finish(&device->instruction_block_pool);
1136    anv_state_pool_finish(&device->dynamic_state_pool);
1137    anv_block_pool_finish(&device->dynamic_state_block_pool);
1138
1139    anv_bo_pool_finish(&device->batch_bo_pool);
1140
1141    pthread_cond_destroy(&device->queue_submit);
1142    pthread_mutex_destroy(&device->mutex);
1143
1144    anv_gem_destroy_context(device, device->context_id);
1145
1146    close(device->fd);
1147
1148    vk_free(&device->alloc, device);
1149 }
1150
1151 VkResult anv_EnumerateInstanceExtensionProperties(
1152     const char*                                 pLayerName,
1153     uint32_t*                                   pPropertyCount,
1154     VkExtensionProperties*                      pProperties)
1155 {
1156    if (pProperties == NULL) {
1157       *pPropertyCount = ARRAY_SIZE(global_extensions);
1158       return VK_SUCCESS;
1159    }
1160
1161    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(global_extensions));
1162    typed_memcpy(pProperties, global_extensions, *pPropertyCount);
1163
1164    if (*pPropertyCount < ARRAY_SIZE(global_extensions))
1165       return VK_INCOMPLETE;
1166
1167    return VK_SUCCESS;
1168 }
1169
1170 VkResult anv_EnumerateDeviceExtensionProperties(
1171     VkPhysicalDevice                            physicalDevice,
1172     const char*                                 pLayerName,
1173     uint32_t*                                   pPropertyCount,
1174     VkExtensionProperties*                      pProperties)
1175 {
1176    if (pProperties == NULL) {
1177       *pPropertyCount = ARRAY_SIZE(device_extensions);
1178       return VK_SUCCESS;
1179    }
1180
1181    *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(device_extensions));
1182    typed_memcpy(pProperties, device_extensions, *pPropertyCount);
1183
1184    if (*pPropertyCount < ARRAY_SIZE(device_extensions))
1185       return VK_INCOMPLETE;
1186
1187    return VK_SUCCESS;
1188 }
1189
1190 VkResult anv_EnumerateInstanceLayerProperties(
1191     uint32_t*                                   pPropertyCount,
1192     VkLayerProperties*                          pProperties)
1193 {
1194    if (pProperties == NULL) {
1195       *pPropertyCount = 0;
1196       return VK_SUCCESS;
1197    }
1198
1199    /* None supported at this time */
1200    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1201 }
1202
1203 VkResult anv_EnumerateDeviceLayerProperties(
1204     VkPhysicalDevice                            physicalDevice,
1205     uint32_t*                                   pPropertyCount,
1206     VkLayerProperties*                          pProperties)
1207 {
1208    if (pProperties == NULL) {
1209       *pPropertyCount = 0;
1210       return VK_SUCCESS;
1211    }
1212
1213    /* None supported at this time */
1214    return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1215 }
1216
1217 void anv_GetDeviceQueue(
1218     VkDevice                                    _device,
1219     uint32_t                                    queueNodeIndex,
1220     uint32_t                                    queueIndex,
1221     VkQueue*                                    pQueue)
1222 {
1223    ANV_FROM_HANDLE(anv_device, device, _device);
1224
1225    assert(queueIndex == 0);
1226
1227    *pQueue = anv_queue_to_handle(&device->queue);
1228 }
1229
1230 VkResult
1231 anv_device_execbuf(struct anv_device *device,
1232                    struct drm_i915_gem_execbuffer2 *execbuf,
1233                    struct anv_bo **execbuf_bos)
1234 {
1235    int ret = anv_gem_execbuffer(device, execbuf);
1236    if (ret != 0) {
1237       /* We don't know the real error. */
1238       return vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
1239    }
1240
1241    struct drm_i915_gem_exec_object2 *objects =
1242       (void *)(uintptr_t)execbuf->buffers_ptr;
1243    for (uint32_t k = 0; k < execbuf->buffer_count; k++)
1244       execbuf_bos[k]->offset = objects[k].offset;
1245
1246    return VK_SUCCESS;
1247 }
1248
1249 VkResult anv_QueueSubmit(
1250     VkQueue                                     _queue,
1251     uint32_t                                    submitCount,
1252     const VkSubmitInfo*                         pSubmits,
1253     VkFence                                     _fence)
1254 {
1255    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1256    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1257    struct anv_device *device = queue->device;
1258    VkResult result = VK_SUCCESS;
1259
1260    /* We lock around QueueSubmit for three main reasons:
1261     *
1262     *  1) When a block pool is resized, we create a new gem handle with a
1263     *     different size and, in the case of surface states, possibly a
1264     *     different center offset but we re-use the same anv_bo struct when
1265     *     we do so.  If this happens in the middle of setting up an execbuf,
1266     *     we could end up with our list of BOs out of sync with our list of
1267     *     gem handles.
1268     *
1269     *  2) The algorithm we use for building the list of unique buffers isn't
1270     *     thread-safe.  While the client is supposed to syncronize around
1271     *     QueueSubmit, this would be extremely difficult to debug if it ever
1272     *     came up in the wild due to a broken app.  It's better to play it
1273     *     safe and just lock around QueueSubmit.
1274     *
1275     *  3)  The anv_cmd_buffer_execbuf function may perform relocations in
1276     *      userspace.  Due to the fact that the surface state buffer is shared
1277     *      between batches, we can't afford to have that happen from multiple
1278     *      threads at the same time.  Even though the user is supposed to
1279     *      ensure this doesn't happen, we play it safe as in (2) above.
1280     *
1281     * Since the only other things that ever take the device lock such as block
1282     * pool resize only rarely happen, this will almost never be contended so
1283     * taking a lock isn't really an expensive operation in this case.
1284     */
1285    pthread_mutex_lock(&device->mutex);
1286
1287    for (uint32_t i = 0; i < submitCount; i++) {
1288       for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
1289          ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
1290                          pSubmits[i].pCommandBuffers[j]);
1291          assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1292
1293          result = anv_cmd_buffer_execbuf(device, cmd_buffer);
1294          if (result != VK_SUCCESS)
1295             goto out;
1296       }
1297    }
1298
1299    if (fence) {
1300       struct anv_bo *fence_bo = &fence->bo;
1301       result = anv_device_execbuf(device, &fence->execbuf, &fence_bo);
1302       if (result != VK_SUCCESS)
1303          goto out;
1304
1305       /* Update the fence and wake up any waiters */
1306       assert(fence->state == ANV_FENCE_STATE_RESET);
1307       fence->state = ANV_FENCE_STATE_SUBMITTED;
1308       pthread_cond_broadcast(&device->queue_submit);
1309    }
1310
1311 out:
1312    pthread_mutex_unlock(&device->mutex);
1313
1314    return result;
1315 }
1316
1317 VkResult anv_QueueWaitIdle(
1318     VkQueue                                     _queue)
1319 {
1320    ANV_FROM_HANDLE(anv_queue, queue, _queue);
1321
1322    return anv_DeviceWaitIdle(anv_device_to_handle(queue->device));
1323 }
1324
1325 VkResult anv_DeviceWaitIdle(
1326     VkDevice                                    _device)
1327 {
1328    ANV_FROM_HANDLE(anv_device, device, _device);
1329    struct anv_batch batch;
1330
1331    uint32_t cmds[8];
1332    batch.start = batch.next = cmds;
1333    batch.end = (void *) cmds + sizeof(cmds);
1334
1335    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1336    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1337
1338    return anv_device_submit_simple_batch(device, &batch);
1339 }
1340
1341 VkResult
1342 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1343 {
1344    uint32_t gem_handle = anv_gem_create(device, size);
1345    if (!gem_handle)
1346       return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1347
1348    anv_bo_init(bo, gem_handle, size);
1349
1350    return VK_SUCCESS;
1351 }
1352
1353 VkResult anv_AllocateMemory(
1354     VkDevice                                    _device,
1355     const VkMemoryAllocateInfo*                 pAllocateInfo,
1356     const VkAllocationCallbacks*                pAllocator,
1357     VkDeviceMemory*                             pMem)
1358 {
1359    ANV_FROM_HANDLE(anv_device, device, _device);
1360    struct anv_device_memory *mem;
1361    VkResult result;
1362
1363    assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1364
1365    /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1366    assert(pAllocateInfo->allocationSize > 0);
1367
1368    /* We support exactly one memory heap. */
1369    assert(pAllocateInfo->memoryTypeIndex == 0 ||
1370           (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1371
1372    /* FINISHME: Fail if allocation request exceeds heap size. */
1373
1374    mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1375                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1376    if (mem == NULL)
1377       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1378
1379    /* The kernel is going to give us whole pages anyway */
1380    uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1381
1382    result = anv_bo_init_new(&mem->bo, device, alloc_size);
1383    if (result != VK_SUCCESS)
1384       goto fail;
1385
1386    mem->type_index = pAllocateInfo->memoryTypeIndex;
1387
1388    mem->map = NULL;
1389    mem->map_size = 0;
1390
1391    *pMem = anv_device_memory_to_handle(mem);
1392
1393    return VK_SUCCESS;
1394
1395  fail:
1396    vk_free2(&device->alloc, pAllocator, mem);
1397
1398    return result;
1399 }
1400
1401 void anv_FreeMemory(
1402     VkDevice                                    _device,
1403     VkDeviceMemory                              _mem,
1404     const VkAllocationCallbacks*                pAllocator)
1405 {
1406    ANV_FROM_HANDLE(anv_device, device, _device);
1407    ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1408
1409    if (mem == NULL)
1410       return;
1411
1412    if (mem->map)
1413       anv_UnmapMemory(_device, _mem);
1414
1415    if (mem->bo.map)
1416       anv_gem_munmap(mem->bo.map, mem->bo.size);
1417
1418    if (mem->bo.gem_handle != 0)
1419       anv_gem_close(device, mem->bo.gem_handle);
1420
1421    vk_free2(&device->alloc, pAllocator, mem);
1422 }
1423
1424 VkResult anv_MapMemory(
1425     VkDevice                                    _device,
1426     VkDeviceMemory                              _memory,
1427     VkDeviceSize                                offset,
1428     VkDeviceSize                                size,
1429     VkMemoryMapFlags                            flags,
1430     void**                                      ppData)
1431 {
1432    ANV_FROM_HANDLE(anv_device, device, _device);
1433    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1434
1435    if (mem == NULL) {
1436       *ppData = NULL;
1437       return VK_SUCCESS;
1438    }
1439
1440    if (size == VK_WHOLE_SIZE)
1441       size = mem->bo.size - offset;
1442
1443    /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1444     *
1445     *  * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1446     *    assert(size != 0);
1447     *  * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1448     *    equal to the size of the memory minus offset
1449     */
1450    assert(size > 0);
1451    assert(offset + size <= mem->bo.size);
1452
1453    /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1454     * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1455     * at a time is valid. We could just mmap up front and return an offset
1456     * pointer here, but that may exhaust virtual memory on 32 bit
1457     * userspace. */
1458
1459    uint32_t gem_flags = 0;
1460    if (!device->info.has_llc && mem->type_index == 0)
1461       gem_flags |= I915_MMAP_WC;
1462
1463    /* GEM will fail to map if the offset isn't 4k-aligned.  Round down. */
1464    uint64_t map_offset = offset & ~4095ull;
1465    assert(offset >= map_offset);
1466    uint64_t map_size = (offset + size) - map_offset;
1467
1468    /* Let's map whole pages */
1469    map_size = align_u64(map_size, 4096);
1470
1471    void *map = anv_gem_mmap(device, mem->bo.gem_handle,
1472                             map_offset, map_size, gem_flags);
1473    if (map == MAP_FAILED)
1474       return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
1475
1476    mem->map = map;
1477    mem->map_size = map_size;
1478
1479    *ppData = mem->map + (offset - map_offset);
1480
1481    return VK_SUCCESS;
1482 }
1483
1484 void anv_UnmapMemory(
1485     VkDevice                                    _device,
1486     VkDeviceMemory                              _memory)
1487 {
1488    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1489
1490    if (mem == NULL)
1491       return;
1492
1493    anv_gem_munmap(mem->map, mem->map_size);
1494
1495    mem->map = NULL;
1496    mem->map_size = 0;
1497 }
1498
1499 static void
1500 clflush_mapped_ranges(struct anv_device         *device,
1501                       uint32_t                   count,
1502                       const VkMappedMemoryRange *ranges)
1503 {
1504    for (uint32_t i = 0; i < count; i++) {
1505       ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1506       if (ranges[i].offset >= mem->map_size)
1507          continue;
1508
1509       anv_clflush_range(mem->map + ranges[i].offset,
1510                         MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
1511    }
1512 }
1513
1514 VkResult anv_FlushMappedMemoryRanges(
1515     VkDevice                                    _device,
1516     uint32_t                                    memoryRangeCount,
1517     const VkMappedMemoryRange*                  pMemoryRanges)
1518 {
1519    ANV_FROM_HANDLE(anv_device, device, _device);
1520
1521    if (device->info.has_llc)
1522       return VK_SUCCESS;
1523
1524    /* Make sure the writes we're flushing have landed. */
1525    __builtin_ia32_mfence();
1526
1527    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1528
1529    return VK_SUCCESS;
1530 }
1531
1532 VkResult anv_InvalidateMappedMemoryRanges(
1533     VkDevice                                    _device,
1534     uint32_t                                    memoryRangeCount,
1535     const VkMappedMemoryRange*                  pMemoryRanges)
1536 {
1537    ANV_FROM_HANDLE(anv_device, device, _device);
1538
1539    if (device->info.has_llc)
1540       return VK_SUCCESS;
1541
1542    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1543
1544    /* Make sure no reads get moved up above the invalidate. */
1545    __builtin_ia32_mfence();
1546
1547    return VK_SUCCESS;
1548 }
1549
1550 void anv_GetBufferMemoryRequirements(
1551     VkDevice                                    _device,
1552     VkBuffer                                    _buffer,
1553     VkMemoryRequirements*                       pMemoryRequirements)
1554 {
1555    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1556    ANV_FROM_HANDLE(anv_device, device, _device);
1557
1558    /* The Vulkan spec (git aaed022) says:
1559     *
1560     *    memoryTypeBits is a bitfield and contains one bit set for every
1561     *    supported memory type for the resource. The bit `1<<i` is set if and
1562     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1563     *    structure for the physical device is supported.
1564     *
1565     * We support exactly one memory type on LLC, two on non-LLC.
1566     */
1567    pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
1568
1569    pMemoryRequirements->size = buffer->size;
1570    pMemoryRequirements->alignment = 16;
1571 }
1572
1573 void anv_GetImageMemoryRequirements(
1574     VkDevice                                    _device,
1575     VkImage                                     _image,
1576     VkMemoryRequirements*                       pMemoryRequirements)
1577 {
1578    ANV_FROM_HANDLE(anv_image, image, _image);
1579    ANV_FROM_HANDLE(anv_device, device, _device);
1580
1581    /* The Vulkan spec (git aaed022) says:
1582     *
1583     *    memoryTypeBits is a bitfield and contains one bit set for every
1584     *    supported memory type for the resource. The bit `1<<i` is set if and
1585     *    only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1586     *    structure for the physical device is supported.
1587     *
1588     * We support exactly one memory type on LLC, two on non-LLC.
1589     */
1590    pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
1591
1592    pMemoryRequirements->size = image->size;
1593    pMemoryRequirements->alignment = image->alignment;
1594 }
1595
1596 void anv_GetImageSparseMemoryRequirements(
1597     VkDevice                                    device,
1598     VkImage                                     image,
1599     uint32_t*                                   pSparseMemoryRequirementCount,
1600     VkSparseImageMemoryRequirements*            pSparseMemoryRequirements)
1601 {
1602    stub();
1603 }
1604
1605 void anv_GetDeviceMemoryCommitment(
1606     VkDevice                                    device,
1607     VkDeviceMemory                              memory,
1608     VkDeviceSize*                               pCommittedMemoryInBytes)
1609 {
1610    *pCommittedMemoryInBytes = 0;
1611 }
1612
1613 VkResult anv_BindBufferMemory(
1614     VkDevice                                    device,
1615     VkBuffer                                    _buffer,
1616     VkDeviceMemory                              _memory,
1617     VkDeviceSize                                memoryOffset)
1618 {
1619    ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1620    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1621
1622    if (mem) {
1623       buffer->bo = &mem->bo;
1624       buffer->offset = memoryOffset;
1625    } else {
1626       buffer->bo = NULL;
1627       buffer->offset = 0;
1628    }
1629
1630    return VK_SUCCESS;
1631 }
1632
1633 VkResult anv_QueueBindSparse(
1634     VkQueue                                     queue,
1635     uint32_t                                    bindInfoCount,
1636     const VkBindSparseInfo*                     pBindInfo,
1637     VkFence                                     fence)
1638 {
1639    stub_return(VK_ERROR_INCOMPATIBLE_DRIVER);
1640 }
1641
1642 VkResult anv_CreateFence(
1643     VkDevice                                    _device,
1644     const VkFenceCreateInfo*                    pCreateInfo,
1645     const VkAllocationCallbacks*                pAllocator,
1646     VkFence*                                    pFence)
1647 {
1648    ANV_FROM_HANDLE(anv_device, device, _device);
1649    struct anv_bo fence_bo;
1650    struct anv_fence *fence;
1651    struct anv_batch batch;
1652    VkResult result;
1653
1654    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1655
1656    result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo, 4096);
1657    if (result != VK_SUCCESS)
1658       return result;
1659
1660    /* Fences are small.  Just store the CPU data structure in the BO. */
1661    fence = fence_bo.map;
1662    fence->bo = fence_bo;
1663
1664    /* Place the batch after the CPU data but on its own cache line. */
1665    const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
1666    batch.next = batch.start = fence->bo.map + batch_offset;
1667    batch.end = fence->bo.map + fence->bo.size;
1668    anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1669    anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1670
1671    if (!device->info.has_llc) {
1672       assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
1673       assert(batch.next - batch.start <= CACHELINE_SIZE);
1674       __builtin_ia32_mfence();
1675       __builtin_ia32_clflush(batch.start);
1676    }
1677
1678    fence->exec2_objects[0].handle = fence->bo.gem_handle;
1679    fence->exec2_objects[0].relocation_count = 0;
1680    fence->exec2_objects[0].relocs_ptr = 0;
1681    fence->exec2_objects[0].alignment = 0;
1682    fence->exec2_objects[0].offset = fence->bo.offset;
1683    fence->exec2_objects[0].flags = 0;
1684    fence->exec2_objects[0].rsvd1 = 0;
1685    fence->exec2_objects[0].rsvd2 = 0;
1686
1687    fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1688    fence->execbuf.buffer_count = 1;
1689    fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
1690    fence->execbuf.batch_len = batch.next - batch.start;
1691    fence->execbuf.cliprects_ptr = 0;
1692    fence->execbuf.num_cliprects = 0;
1693    fence->execbuf.DR1 = 0;
1694    fence->execbuf.DR4 = 0;
1695
1696    fence->execbuf.flags =
1697       I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1698    fence->execbuf.rsvd1 = device->context_id;
1699    fence->execbuf.rsvd2 = 0;
1700
1701    if (pCreateInfo->flags & VK_FENCE_CREATE_SIGNALED_BIT) {
1702       fence->state = ANV_FENCE_STATE_SIGNALED;
1703    } else {
1704       fence->state = ANV_FENCE_STATE_RESET;
1705    }
1706
1707    *pFence = anv_fence_to_handle(fence);
1708
1709    return VK_SUCCESS;
1710 }
1711
1712 void anv_DestroyFence(
1713     VkDevice                                    _device,
1714     VkFence                                     _fence,
1715     const VkAllocationCallbacks*                pAllocator)
1716 {
1717    ANV_FROM_HANDLE(anv_device, device, _device);
1718    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1719
1720    if (!fence)
1721       return;
1722
1723    assert(fence->bo.map == fence);
1724    anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
1725 }
1726
1727 VkResult anv_ResetFences(
1728     VkDevice                                    _device,
1729     uint32_t                                    fenceCount,
1730     const VkFence*                              pFences)
1731 {
1732    for (uint32_t i = 0; i < fenceCount; i++) {
1733       ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1734       fence->state = ANV_FENCE_STATE_RESET;
1735    }
1736
1737    return VK_SUCCESS;
1738 }
1739
1740 VkResult anv_GetFenceStatus(
1741     VkDevice                                    _device,
1742     VkFence                                     _fence)
1743 {
1744    ANV_FROM_HANDLE(anv_device, device, _device);
1745    ANV_FROM_HANDLE(anv_fence, fence, _fence);
1746    int64_t t = 0;
1747    int ret;
1748
1749    switch (fence->state) {
1750    case ANV_FENCE_STATE_RESET:
1751       /* If it hasn't even been sent off to the GPU yet, it's not ready */
1752       return VK_NOT_READY;
1753
1754    case ANV_FENCE_STATE_SIGNALED:
1755       /* It's been signaled, return success */
1756       return VK_SUCCESS;
1757
1758    case ANV_FENCE_STATE_SUBMITTED:
1759       /* It's been submitted to the GPU but we don't know if it's done yet. */
1760       ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1761       if (ret == 0) {
1762          fence->state = ANV_FENCE_STATE_SIGNALED;
1763          return VK_SUCCESS;
1764       } else {
1765          return VK_NOT_READY;
1766       }
1767    default:
1768       unreachable("Invalid fence status");
1769    }
1770 }
1771
1772 #define NSEC_PER_SEC 1000000000
1773 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
1774
1775 VkResult anv_WaitForFences(
1776     VkDevice                                    _device,
1777     uint32_t                                    fenceCount,
1778     const VkFence*                              pFences,
1779     VkBool32                                    waitAll,
1780     uint64_t                                    _timeout)
1781 {
1782    ANV_FROM_HANDLE(anv_device, device, _device);
1783    int ret;
1784
1785    /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1786     * to block indefinitely timeouts <= 0.  Unfortunately, this was broken
1787     * for a couple of kernel releases.  Since there's no way to know
1788     * whether or not the kernel we're using is one of the broken ones, the
1789     * best we can do is to clamp the timeout to INT64_MAX.  This limits the
1790     * maximum timeout from 584 years to 292 years - likely not a big deal.
1791     */
1792    int64_t timeout = MIN2(_timeout, INT64_MAX);
1793
1794    uint32_t pending_fences = fenceCount;
1795    while (pending_fences) {
1796       pending_fences = 0;
1797       bool signaled_fences = false;
1798       for (uint32_t i = 0; i < fenceCount; i++) {
1799          ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1800          switch (fence->state) {
1801          case ANV_FENCE_STATE_RESET:
1802             /* This fence hasn't been submitted yet, we'll catch it the next
1803              * time around.  Yes, this may mean we dead-loop but, short of
1804              * lots of locking and a condition variable, there's not much that
1805              * we can do about that.
1806              */
1807             pending_fences++;
1808             continue;
1809
1810          case ANV_FENCE_STATE_SIGNALED:
1811             /* This fence is not pending.  If waitAll isn't set, we can return
1812              * early.  Otherwise, we have to keep going.
1813              */
1814             if (!waitAll)
1815                return VK_SUCCESS;
1816             continue;
1817
1818          case ANV_FENCE_STATE_SUBMITTED:
1819             /* These are the fences we really care about.  Go ahead and wait
1820              * on it until we hit a timeout.
1821              */
1822             ret = anv_gem_wait(device, fence->bo.gem_handle, &timeout);
1823             if (ret == -1 && errno == ETIME) {
1824                return VK_TIMEOUT;
1825             } else if (ret == -1) {
1826                /* We don't know the real error. */
1827                return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1828             } else {
1829                fence->state = ANV_FENCE_STATE_SIGNALED;
1830                signaled_fences = true;
1831                if (!waitAll)
1832                   return VK_SUCCESS;
1833                continue;
1834             }
1835          }
1836       }
1837
1838       if (pending_fences && !signaled_fences) {
1839          /* If we've hit this then someone decided to vkWaitForFences before
1840           * they've actually submitted any of them to a queue.  This is a
1841           * fairly pessimal case, so it's ok to lock here and use a standard
1842           * pthreads condition variable.
1843           */
1844          pthread_mutex_lock(&device->mutex);
1845
1846          /* It's possible that some of the fences have changed state since the
1847           * last time we checked.  Now that we have the lock, check for
1848           * pending fences again and don't wait if it's changed.
1849           */
1850          uint32_t now_pending_fences = 0;
1851          for (uint32_t i = 0; i < fenceCount; i++) {
1852             ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1853             if (fence->state == ANV_FENCE_STATE_RESET)
1854                now_pending_fences++;
1855          }
1856          assert(now_pending_fences <= pending_fences);
1857
1858          if (now_pending_fences == pending_fences) {
1859             struct timespec before;
1860             clock_gettime(CLOCK_MONOTONIC, &before);
1861
1862             uint32_t abs_nsec = before.tv_nsec + timeout % NSEC_PER_SEC;
1863             uint64_t abs_sec = before.tv_sec + (abs_nsec / NSEC_PER_SEC) +
1864                                (timeout / NSEC_PER_SEC);
1865             abs_nsec %= NSEC_PER_SEC;
1866
1867             /* Avoid roll-over in tv_sec on 32-bit systems if the user
1868              * provided timeout is UINT64_MAX
1869              */
1870             struct timespec abstime;
1871             abstime.tv_nsec = abs_nsec;
1872             abstime.tv_sec = MIN2(abs_sec, INT_TYPE_MAX(abstime.tv_sec));
1873
1874             ret = pthread_cond_timedwait(&device->queue_submit,
1875                                          &device->mutex, &abstime);
1876             assert(ret != EINVAL);
1877
1878             struct timespec after;
1879             clock_gettime(CLOCK_MONOTONIC, &after);
1880             uint64_t time_elapsed =
1881                ((uint64_t)after.tv_sec * NSEC_PER_SEC + after.tv_nsec) -
1882                ((uint64_t)before.tv_sec * NSEC_PER_SEC + before.tv_nsec);
1883
1884             if (time_elapsed >= timeout) {
1885                pthread_mutex_unlock(&device->mutex);
1886                return VK_TIMEOUT;
1887             }
1888
1889             timeout -= time_elapsed;
1890          }
1891
1892          pthread_mutex_unlock(&device->mutex);
1893       }
1894    }
1895
1896    return VK_SUCCESS;
1897 }
1898
1899 // Queue semaphore functions
1900
1901 VkResult anv_CreateSemaphore(
1902     VkDevice                                    device,
1903     const VkSemaphoreCreateInfo*                pCreateInfo,
1904     const VkAllocationCallbacks*                pAllocator,
1905     VkSemaphore*                                pSemaphore)
1906 {
1907    /* The DRM execbuffer ioctl always execute in-oder, even between different
1908     * rings. As such, there's nothing to do for the user space semaphore.
1909     */
1910
1911    *pSemaphore = (VkSemaphore)1;
1912
1913    return VK_SUCCESS;
1914 }
1915
1916 void anv_DestroySemaphore(
1917     VkDevice                                    device,
1918     VkSemaphore                                 semaphore,
1919     const VkAllocationCallbacks*                pAllocator)
1920 {
1921 }
1922
1923 // Event functions
1924
1925 VkResult anv_CreateEvent(
1926     VkDevice                                    _device,
1927     const VkEventCreateInfo*                    pCreateInfo,
1928     const VkAllocationCallbacks*                pAllocator,
1929     VkEvent*                                    pEvent)
1930 {
1931    ANV_FROM_HANDLE(anv_device, device, _device);
1932    struct anv_state state;
1933    struct anv_event *event;
1934
1935    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1936
1937    state = anv_state_pool_alloc(&device->dynamic_state_pool,
1938                                 sizeof(*event), 8);
1939    event = state.map;
1940    event->state = state;
1941    event->semaphore = VK_EVENT_RESET;
1942
1943    if (!device->info.has_llc) {
1944       /* Make sure the writes we're flushing have landed. */
1945       __builtin_ia32_mfence();
1946       __builtin_ia32_clflush(event);
1947    }
1948
1949    *pEvent = anv_event_to_handle(event);
1950
1951    return VK_SUCCESS;
1952 }
1953
1954 void anv_DestroyEvent(
1955     VkDevice                                    _device,
1956     VkEvent                                     _event,
1957     const VkAllocationCallbacks*                pAllocator)
1958 {
1959    ANV_FROM_HANDLE(anv_device, device, _device);
1960    ANV_FROM_HANDLE(anv_event, event, _event);
1961
1962    if (!event)
1963       return;
1964
1965    anv_state_pool_free(&device->dynamic_state_pool, event->state);
1966 }
1967
1968 VkResult anv_GetEventStatus(
1969     VkDevice                                    _device,
1970     VkEvent                                     _event)
1971 {
1972    ANV_FROM_HANDLE(anv_device, device, _device);
1973    ANV_FROM_HANDLE(anv_event, event, _event);
1974
1975    if (!device->info.has_llc) {
1976       /* Invalidate read cache before reading event written by GPU. */
1977       __builtin_ia32_clflush(event);
1978       __builtin_ia32_mfence();
1979
1980    }
1981
1982    return event->semaphore;
1983 }
1984
1985 VkResult anv_SetEvent(
1986     VkDevice                                    _device,
1987     VkEvent                                     _event)
1988 {
1989    ANV_FROM_HANDLE(anv_device, device, _device);
1990    ANV_FROM_HANDLE(anv_event, event, _event);
1991
1992    event->semaphore = VK_EVENT_SET;
1993
1994    if (!device->info.has_llc) {
1995       /* Make sure the writes we're flushing have landed. */
1996       __builtin_ia32_mfence();
1997       __builtin_ia32_clflush(event);
1998    }
1999
2000    return VK_SUCCESS;
2001 }
2002
2003 VkResult anv_ResetEvent(
2004     VkDevice                                    _device,
2005     VkEvent                                     _event)
2006 {
2007    ANV_FROM_HANDLE(anv_device, device, _device);
2008    ANV_FROM_HANDLE(anv_event, event, _event);
2009
2010    event->semaphore = VK_EVENT_RESET;
2011
2012    if (!device->info.has_llc) {
2013       /* Make sure the writes we're flushing have landed. */
2014       __builtin_ia32_mfence();
2015       __builtin_ia32_clflush(event);
2016    }
2017
2018    return VK_SUCCESS;
2019 }
2020
2021 // Buffer functions
2022
2023 VkResult anv_CreateBuffer(
2024     VkDevice                                    _device,
2025     const VkBufferCreateInfo*                   pCreateInfo,
2026     const VkAllocationCallbacks*                pAllocator,
2027     VkBuffer*                                   pBuffer)
2028 {
2029    ANV_FROM_HANDLE(anv_device, device, _device);
2030    struct anv_buffer *buffer;
2031
2032    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
2033
2034    buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
2035                        VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2036    if (buffer == NULL)
2037       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2038
2039    buffer->size = pCreateInfo->size;
2040    buffer->usage = pCreateInfo->usage;
2041    buffer->bo = NULL;
2042    buffer->offset = 0;
2043
2044    *pBuffer = anv_buffer_to_handle(buffer);
2045
2046    return VK_SUCCESS;
2047 }
2048
2049 void anv_DestroyBuffer(
2050     VkDevice                                    _device,
2051     VkBuffer                                    _buffer,
2052     const VkAllocationCallbacks*                pAllocator)
2053 {
2054    ANV_FROM_HANDLE(anv_device, device, _device);
2055    ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
2056
2057    if (!buffer)
2058       return;
2059
2060    vk_free2(&device->alloc, pAllocator, buffer);
2061 }
2062
2063 void
2064 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
2065                               enum isl_format format,
2066                               uint32_t offset, uint32_t range, uint32_t stride)
2067 {
2068    isl_buffer_fill_state(&device->isl_dev, state.map,
2069                          .address = offset,
2070                          .mocs = device->default_mocs,
2071                          .size = range,
2072                          .format = format,
2073                          .stride = stride);
2074
2075    anv_state_flush(device, state);
2076 }
2077
2078 void anv_DestroySampler(
2079     VkDevice                                    _device,
2080     VkSampler                                   _sampler,
2081     const VkAllocationCallbacks*                pAllocator)
2082 {
2083    ANV_FROM_HANDLE(anv_device, device, _device);
2084    ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
2085
2086    if (!sampler)
2087       return;
2088
2089    vk_free2(&device->alloc, pAllocator, sampler);
2090 }
2091
2092 VkResult anv_CreateFramebuffer(
2093     VkDevice                                    _device,
2094     const VkFramebufferCreateInfo*              pCreateInfo,
2095     const VkAllocationCallbacks*                pAllocator,
2096     VkFramebuffer*                              pFramebuffer)
2097 {
2098    ANV_FROM_HANDLE(anv_device, device, _device);
2099    struct anv_framebuffer *framebuffer;
2100
2101    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2102
2103    size_t size = sizeof(*framebuffer) +
2104                  sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
2105    framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2106                             VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2107    if (framebuffer == NULL)
2108       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2109
2110    framebuffer->attachment_count = pCreateInfo->attachmentCount;
2111    for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2112       VkImageView _iview = pCreateInfo->pAttachments[i];
2113       framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
2114    }
2115
2116    framebuffer->width = pCreateInfo->width;
2117    framebuffer->height = pCreateInfo->height;
2118    framebuffer->layers = pCreateInfo->layers;
2119
2120    *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
2121
2122    return VK_SUCCESS;
2123 }
2124
2125 void anv_DestroyFramebuffer(
2126     VkDevice                                    _device,
2127     VkFramebuffer                               _fb,
2128     const VkAllocationCallbacks*                pAllocator)
2129 {
2130    ANV_FROM_HANDLE(anv_device, device, _device);
2131    ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
2132
2133    if (!fb)
2134       return;
2135
2136    vk_free2(&device->alloc, pAllocator, fb);
2137 }
2138
2139 /* vk_icd.h does not declare this function, so we declare it here to
2140  * suppress Wmissing-prototypes.
2141  */
2142 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2143 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
2144
2145 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2146 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
2147 {
2148    /* For the full details on loader interface versioning, see
2149     * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2150     * What follows is a condensed summary, to help you navigate the large and
2151     * confusing official doc.
2152     *
2153     *   - Loader interface v0 is incompatible with later versions. We don't
2154     *     support it.
2155     *
2156     *   - In loader interface v1:
2157     *       - The first ICD entrypoint called by the loader is
2158     *         vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2159     *         entrypoint.
2160     *       - The ICD must statically expose no other Vulkan symbol unless it is
2161     *         linked with -Bsymbolic.
2162     *       - Each dispatchable Vulkan handle created by the ICD must be
2163     *         a pointer to a struct whose first member is VK_LOADER_DATA. The
2164     *         ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2165     *       - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2166     *         vkDestroySurfaceKHR(). The ICD must be capable of working with
2167     *         such loader-managed surfaces.
2168     *
2169     *    - Loader interface v2 differs from v1 in:
2170     *       - The first ICD entrypoint called by the loader is
2171     *         vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2172     *         statically expose this entrypoint.
2173     *
2174     *    - Loader interface v3 differs from v2 in:
2175     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2176     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2177     *          because the loader no longer does so.
2178     */
2179    *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2180    return VK_SUCCESS;
2181 }